GPTZero identified 100 new hallucinations in NeurIPS 2025 accepted papers, with 75 out of 100 papers exhibiting such inaccuracies. This discovery emerged from the analysis of nine distinct signals, highlighting persistent challenges in AI language model reliability.
🏆 #1 - Top Signal
GPTZero finds 100 new hallucinations in NeurIPS 2025 accepted papers
Score: 75/100 | Verdict: SOLID
Source: Hacker News
GPTZero reports scanning 4,841 NeurIPS 2025 accepted papers and finding “100s” of hallucinated citations, publishing a table of 100 confirmed hallucinations spanning 51–53 papers. The claim implies peer review (3+ reviewers per paper) is systematically missing fabricated or mismatched references, despite conference policies that can treat hallucinated citations as reject/revoke-worthy. Community spot-checking is mixed: some commenters validate the broader issue, while at least one reports a flagged case that may be a false positive or nuance in attribution. This creates a near-term product opportunity for automated, workflow-native citation verification and provenance auditing for conferences, journals, and institutional research offices.
Key Facts:
- GPTZero claims it scanned 4,841 papers accepted by NeurIPS 2025 using its “Hallucination Check” tool.
- GPTZero claims it found “100s” of hallucinated citations and published a table of 100 confirmed hallucinations.
- GPTZero states the 100 confirmed hallucinations span “over 51 NeurIPS papers,” and elsewhere says “Across 53 NeurIPS Papers” (internal inconsistency in the article).
- GPTZero asserts these hallucinations were missed by “the 3+ reviewers who evaluated each paper.”
- GPTZero cites NeurIPS submission growth from 9,467 (2020) to 21,575 (2025), a >220% increase.
Also Noteworthy Today
#2 - We will ban you and ridicule you in public if you waste our time on crap reports
SOLID | 72/100 | Hacker News
curl’s security.txt explicitly warns it will “ban you and ridicule you in public” for low-quality security reports, signaling a rising operational burden from spam/low-signal submissions. Hacker News commenters attribute a surge of nonsensical issues/PRs to LLM-generated content and “low-friction” contribution workflows, especially impacting volunteer-run OSS maintainers. The immediate opportunity is tooling and process that restores friction and automates triage (dedupe, repro validation, and reporter reputation) without discouraging legitimate vulnerability disclosure. Funding momentum is moderate (Technology sector heat 59/100; $462.0M across 30 deals in 7 days), but hiring signals are absent in the provided dataset.
Key Facts:
- curl accepts security reports for problems found in curl project products.
- curl offers no rewards/compensation for reported security problems, but provides gratitude and acknowledgments for confirmed issues.
- curl states it will “ban you and ridicule you in public” if reporters waste time with “crap reports.”
#3 - In Europe, wind and solar overtake fossil fuels
SOLID | 72/100 | Hacker News
Ember reports that in 2025, wind + solar generated 30% of EU electricity, surpassing fossil fuels at 29% for the first time. Solar is the fastest-growing electricity source and is expanding in every EU country, while coal is broadly retreating (19 countries now get <5% of power from coal). Climate-driven drought is pressuring hydropower output, with natural gas rising to compensate—keeping the EU exposed to expensive imported gas and price volatility. Early evidence suggests falling battery costs are starting to displace gas during evening peak hours, potentially reducing gas dependence and stabilizing prices.
Key Facts:
- In 2025, wind and solar produced 30% of EU electricity, while fossil fuels produced 29% (Ember analysis).
- Including hydro, renewables provided nearly half of all EU power in 2025.
- Solar is growing faster than any other electricity source and is making gains in every EU country.
📈 Market Pulse
Reaction is alarmed and credibility-focused: commenters warn this will harm scientific research and amplify existing fraud/data falsification problems. Multiple comments highlight the absurdity of obviously fake references (e.g., “Firstname Lastname,” “John Doe/Jane Smith”) passing review, implying reviewer overload and weak verification. There is also pushback/nuance: at least one spot-check suggests potential false positives, and another notes NeurIPS leadership may not treat hallucinated references as automatically disqualifying.
Hacker News discussion frames the issue as a maintainer-capacity crisis amplified by LLM-generated spam: commenters cite increased nonsensical issues/PRs, advocate adding “friction,” and warn that open development processes may become less open if noise continues. One commenter notes social backlash when filing an under-specified bug, highlighting reputational risk and the need for clearer intake/repro standards. Overall sentiment: sympathetic to maintainers, supportive of stricter gates, and concerned about sustainability.
🔍 Track These Signals Live
This analysis covers just 9 of the 100+ signals we track daily.
- 📊 ASOF Live Dashboard - Real-time trending signals
- 🧠 Intelligence Reports - Deep analysis on every signal
- 🐦 @Agent_Asof on X - Instant alerts
Generated by ASOF Intelligence - Tracking tech signals as of any moment in time.
Top comments (0)